3
Wednesday
51 items
- model
Demis Hassabis: Gemma 4 surpasses 150M downloads as new 12B model launches
Demis Hassabis celebrates Gemma 4 reaching over 150 million downloads and announces the release of the new Gemma 4 12B model, which is powerful yet compact enough to run locally on a laptop with 16GB VRAM under an Apache
- industry
Claude: Claude talks building, creativity, and silly ideas with Ben James
Claude sat down with Ben James to discuss creativity, building projects, and the value of silly ideas.
- product
TestingCatalog: Google Dreambeans experiment launches for AI Ultra users in Labs
Google introduced a new Dreambeans experiment in Google Labs that uses Personal Intelligence to deliver daily stories to US-based AI Ultra users on the waitlist.
- model
TestingCatalog: Ideogram 4.0 debuts as SOTA open image generation model
Ideogram released version 4.0 of its open image generation model, which ranks eighth on LM Arena and supports 2K resolution and precise text rendering.
- model
TestingCatalog: Google releases Gemma 4 12B open-source multimodal model
Google's new Gemma 4 12B model is available on Huggingface under Apache 2.0, offering encoder-free multimodal capabilities ideal for local consumer device deployment.
- product
TestingCatalog: Perplexity Personal Computer launches on Windows for Max users
Perplexity has rolled out its Personal Computer application to Max and Enterprise Max users on Windows via a waitlist.
- product
Min Choi: Creator makes full music video with Dreamina AI Seedance 2.0
Min Choi created an entire music video sequence using Dreamina AI and Octo with Seedance 2.0, highlighting its ability to sync with creative workflows.
- product
TestingCatalog: Capafy launches five e-commerce AI skills for store operators
Capafy released five pre-built e-commerce agent skills including video ad makers and listing generators created by experienced operators.
- product
Min Choi: OpenAI launches Codex Sites for instant interactive apps
OpenAI released Codex Sites, allowing users to convert plans, dashboards, and ideas into interactive web apps with a shareable URL.
- policy
Sam Altman: Altman praises new AI EO for balancing safety and US leadership
Sam Altman endorses a new executive order on AI, saying it strikes the right balance between advancing US model leadership, ensuring safety, and equipping trusted defenders with cyber tools.
- policy
Sam Altman: Altman backs new US AI executive order for safety and cyber defense
Altman argues the US should lead on AI by developing top models safely and equipping trusted defenders with cyber tools, praising the new executive order for striking the right balance.
- guideMastering Tool Use and Function Calling in LLMs
Explore function calling and tool use in LLMs, their mechanics, and designing effective tools for AI agents.
Read - generalOpenAI Launches Codex Sites to Turn Documents Into Interactive Apps, Per X Post
According to a post by Min Choi on X, OpenAI has released Codex Sites, a tool that converts plans, dashboards, launch documents, or ideas into interactive apps with URLs. The thread highlights five example use cases: public equity investing, product design, sales, data analytics, and creative production.
Read - guideEnduring Prompt Engineering Strategies for Instruction-Tuned Models
Explore prompt engineering fundamentals that still work for today's instruction-tuned models and understand what changed in AI interactions.
Read - guideUnderstanding LLM Context Windows: Costs and Key Insights
Explore what long context LLMs provide, their quadratic costs, and effective usage strategies for optimal performance.
Read - guideUnderstanding Mixture-of-Experts (MoE): Efficient Scaling of AI Models
Explore how the mixture of experts architecture efficiently scales parameters while minimizing cost per token in AI models.
Read - guideUnderstanding Transformer Attention: The Key to Modern LLMs
Explore how self-attention and transformer architecture drive the performance of LLMs, including insights on scaling and efficiency.
Read - guideUnderstanding Retrieval-Augmented Generation: A Practical Guide
Explore retrieval augmented generation, its mechanics, and advantages over fine-tuning in natural language processing.
Read - productMicrosoft launches MAI models, Copilot super app and OpenClaw Windows agent
Microsoft unveiled a wave of AI products including new MAI-branded foundation models, a Copilot super app with long-running Autopilot agents, and a built-in OpenClaw Companion agent for Windows.
Read - labsOpenAI introduces role-specific Codex plugins, annotations, and Sites preview
OpenAI introduced role-specific plugins, annotations, and a preview of shareable interactive Sites for Codex, reporting that over 5 million people now use the tool weekly. Non-developers represent roughly 20% of users and are growing more than three times as fast as developers, according to the company.
Read - labsOpenAI Calls for International Youth AI Safety Institute Ahead of G7 Summit
OpenAI has called for the creation of an international youth safety institute to advance global standards for age-appropriate AI use ahead of the G7 Leaders' Summit in France. The company outlined nine principles for youth AI safety and detailed existing ChatGPT safeguards for minors.
Read - labsOpenAI frontier models and Codex are now available on AWS
OpenAI has made its frontier models and Codex generally available on AWS via Amazon Bedrock, allowing enterprises to deploy AI within existing security, governance, and procurement workflows. The announcement includes future plans to bring OpenAI's Daybreak cyber capabilities to AWS.
Read - labsOpenAI launches Rosalind Biodefense and expands GPT-Rosalind access to government partners
OpenAI announced the Rosalind Biodefense program to equip trusted developers with GPT-Rosalind for building defensive biosecurity tools. The company is also expanding access to the model for select U.S. and allied government partners, according to the OpenAI Blog.
Read - labsA shared playbook for trustworthy third party evaluations
OpenAI published recommendations for designing trustworthy third-party evaluations of frontier AI models, emphasizing that the surrounding "harness"—the environment, tools, and setup enabling agentic execution—fundamentally shapes measured capabilities and safeguard robustness. The post categorizes evaluation claims and urges evaluators to transparently report their setup, budget, and validity checks to avoid under-elicitation or miscalibrated results.
Read - labsOpenAI Publishes Frontier Governance Framework to Align Safety Practices with Emerging Regulations
OpenAI has published a Frontier Governance Framework detailing how its safety and security practices align with emerging legal requirements such as California’s Transparency in Frontier AI Act and the EU AI Act. The document translates aspects of the company’s internal Preparedness Framework into public governance commitments covering risk assessment, mitigation, and reporting for advanced AI systems.
Read - labsAnthropic releases Claude Opus 4.8 with benchmark gains and new collaboration features
Anthropic announced Claude Opus 4.8, an upgrade to its flagship model that it says improves benchmarks, honesty, and collaboration while maintaining the same price. The release also introduces effort controls for claude.ai, dynamic workflows in Claude Code, and cheaper fast-mode pricing, according to Anthropic News.
Read - labsAnthropic Launches Claude Design, a Visual Collaboration Tool in Research Preview
Anthropic launched Claude Design, a research-preview product powered by Claude Opus 4.7 that lets Claude Pro, Max, Team, and Enterprise subscribers collaborate with Claude to create designs, prototypes, slides, and other visual work. Enterprise organizations require administrator activation to enable access.
Read - labsAnthropic Launches Project Glasswing with Major Tech and Finance Partners to Defensively Deploy AI Cybersecurity Model
Anthropic announced Project Glasswing, a coalition including AWS, Apple, Google, Microsoft, and others, to use its unreleased Claude Mythos Preview model for defensive cybersecurity. The initiative aims to address the dual-use risk of advanced AI vulnerability-discovery capabilities by finding and fixing flaws in critical software before malicious actors can exploit them.
Read - labsAnthropic publishes findings from 80,508-person global AI interviewer study
Anthropic interviewed 80,508 Claude users across 159 countries and 70 languages using an AI interviewer to understand their aspirations and concerns for artificial intelligence. The study found that respondents most often want AI to support professional excellence, personal transformation, and life management, while simultaneously holding multiple fears about the technology's impact.
Read - labsAnthropic Commits to Keeping Claude Ad-Free
Anthropic announced that its AI assistant Claude will remain free of advertising and advertiser influence, funded instead by enterprise contracts and paid subscriptions. The company argues that introducing ads would compromise Claude's role as a trusted space for sensitive conversations, deep work, and genuine assistance.
Read - labsAnthropic Expands Project Glasswing to Approximately 150 New Organizations
Anthropic is expanding Project Glasswing from roughly 50 to approximately 150 partner organizations across more than 15 countries to scan critical codebases for vulnerabilities using Claude Mythos Preview. The company also released Claude Security for public use and cautioned that other AI developers could deploy comparable cyber-capable models without safeguards within 6 to 12 months.
Read - labsGoogle DeepMind Blog Lists New Gemini Models, Scientific AI Tools, and Global Partnerships
The Google DeepMind blog highlights recent posts announcing updates to the Gemini and Gemma model families, AI systems for scientific discovery and weather prediction, and new international partnerships focused on safety and research.
Read - open-sourceHugging Face Blog Community Section Lists Recent AI Articles on Optimization, Safety, and Applications
The Hugging Face blog's community articles page features new posts from contributors covering model optimization, AI safety research, training methodologies, and experimental machine learning projects. Source: Hugging Face (https://huggingface.co/blog).
Read - labsMistral AI Announces Mistral Medium 3
Mistral AI has announced Mistral Medium 3, a new language model that the company says balances state-of-the-art performance with 8× lower costs and simpler enterprise deployability. The model is available today via Mistral La Plateforme and Amazon SageMaker, with additional platform support coming soon.
Read - platformsAlibaba's HappyHorse 1.0 Video Generation Model Available on Replicate
Alibaba's HappyHorse 1.0 video generation model is now accessible via Replicate's API, supporting both text-to-video and image-to-video generation at 720p and 1080p resolutions with durations from 3 to 15 seconds.
Read - platformsReplicate Weekly Bulletin Spotlights FLUX.1 Tools, Open-Source Deepfake Project, and Sleep Research
Replicate's August 23 weekly bulletin highlights new FLUX.1 interfaces for multi-model and multi-LoRA image generation, the open-source Deep-Live-Cam real-time deepfake tool, and a bioRxiv preprint proposing that REM sleep functions as a form of neural synthetic data generation. Entrepreneur Pieter Levels also mentioned Replicate during a recent Lex Fridman podcast appearance, according to the company blog.
Read - platformsReplicate Intelligence #11: Fine-tune FLUX.1, Tavus digital twins, and new AI video and 3D tools
Replicate now lets users fine-tune the FLUX.1 image model with custom images, while its latest weekly bulletin also covers Tavus’s real-time conversational video API, Sketch2Scene’s sketch-to-3D-game pipeline, and Puppet-Master’s object controls for Stable Video Diffusion.
Read - platformsReplicate Intelligence #10: FLUX.1 image-to-image, Streamlit tutorial, and Odyssey agents
Replicate's weekly bulletin reports nearly 5 million first-week predictions for FLUX.1 [schnell] and new image-to-image support for FLUX.1 [dev]. The update also highlights a Streamlit tutorial featuring Replicate's Zeke and Odyssey, a Minecraft agent framework using fine-tuned LLaMA-3.
Read - platformsReplicate Intelligence #9: FLUX.1, SAM 2, Gemma 2 2B, and new AI tools
Replicate's weekly bulletin covers new open-source AI releases including Black Forest Labs' FLUX.1 image generator, Meta's SAM 2 segmentation model, and Google's Gemma 2 2B language model, alongside tools for model interpretability and distributed training research.
Read - platformsReplicate Intelligence #8: Meta Releases Llama 3.1 405B, Mistral Unveils Large 2, and Meta Open-Sources AI Agent Toolkit
Replicate's weekly bulletin covers Meta's release of the Llama 3.1 model family including the 405B parameter model, Mistral AI's Large 2 under a research license, and Meta's open-sourced toolkit for building AI agents. The issue also highlights Meta's PromptGuard for detecting malicious prompts and a new Replicate API endpoint for searching public models.
Read - labsStability AI Releases Stable Audio 3.0 Open-Weight Model Family
Stability AI released Stable Audio 3.0, a family of open-weight audio models trained on fully licensed data, positioning it as a foundation for the audio community's future artistic experimentation.
Read - labsStability AI releases Stable Audio 3.0 open-weights music model family
Stability AI has released Stable Audio 3.0, a family of open-weights music generation models trained on fully licensed data. The suite includes three downloadable models and an enterprise API tier, offering variable-length generation up to six minutes and commercial-use rights under a community license.
Read - labsStability AI Launches Brand Studio, an AI Creative Production Platform for Enterprise Teams
Stability AI has introduced Brand Studio, an end-to-end creative production platform that lets professional teams generate on-brand content using custom AI models, automated workflows, and precision editing tools. The platform offers Core and Enterprise tiers with features like Brand Central, Producer Mode, and Curated Model Routing.
Read - labsStability AI Joins the Tech Coalition
Stability AI announced it has joined the Tech Coalition, a global alliance of technology companies working to combat online child sexual exploitation and abuse. The company previously participated in the coalition's Pathways program in 2025.
Read - labsUniversal Music Group and Stability AI Announce Strategic Alliance for AI Music Tools
Universal Music Group and Stability AI announced a strategic alliance to co-develop professional, fully licensed AI music creation tools. The partnership will center artists in the development process to guide the creation of commercially safe products built on responsibly trained models.
Read - productGoogle’s Nano Banana 2 and Pro hit general availability with video input
Google has moved Nano Banana 2 and Nano Banana Pro to general availability on its APIs, adding video file support for Nano Banana 2.
Read - productElevenLabs unveils Dubbing v2 Alpha with emotional tone preservation
ElevenLabs has launched a new Dubbing v2 Alpha model that translates speech across all languages while retaining the original emotional tone.
Read - productAnthropic launches Dynamic Workflows for Claude Code with ultracode effort level
Anthropic has introduced Dynamic Workflows in Claude Code, enabling reusable, multi-step agent tasks across enterprise plans and APIs, alongside a new "ultracode" effort level that lets Claude autonomously decide when to spawn workflows.
Read - milestoneOpenAI Foundation pledges $250M to AI prosperity and transition research
Sam Altman announced the OpenAI Foundation is committing $250 million to measurement, transition support, and new economic approaches aimed at broadly shared prosperity from AI.
Read - labsOpenAI Blog Lists Recent Updates on Governance, Codex, ChatGPT, and Research
The OpenAI Blog features recent posts covering a frontier governance framework, self-improving tax agents and enterprise Codex deployments, a Gartner leadership recognition for coding agents, and a model-disproven conjecture in discrete geometry. Additional updates include content provenance efforts, a Dell Technologies partnership, new ChatGPT personal finance features, and safety improvements for sensitive conversations.
Read - labsMistral Updates Vibe Agent, Launches Search Toolkit and Physics AI Models
Mistral announced several updates including Vibe agent Work and Code modes with a VS Code extension, a Search Toolkit for production pipelines, and Physics AI models for engineering acceleration. The company also introduced Mistral Medium 3.5 for remote coding agents and revealed plans for the AI Now Summit 2026.
Read